Upload
others
View
3
Download
0
Embed Size (px)
Citation preview
Final Report for Short-Stay Study Abroad Program 2017
School and Dept. School of Multidisciplinary Sciences, Department of Informatics
3 years
Name Trung-Nghia LE
Destination country The United States of America
Receiving univ./institute The University of Dayton
Period of study 29 days
Date of report 11/14/1017 * A final report must be submitted within one month
upon your return.
University of Dayton
The University of Dayton (UD) is an American private Roman Catholic national research
university in Dayton city, Ohio state. Founded in 1850 by the Society of Mary (Marianists), it is
one of three Marianist universities in the nation and the second-largest private university in Ohio.
The University has about 8,000 undergraduate and 2,200 post-graduate students from a variety of
religious, ethnic and geographic backgrounds, drawn from across the United States and more than
40 countries. It offers more than 80 academic programs in arts and sciences, business
administration, education and health sciences, engineering, law and, in 1988, was first in the
country to offer an undergraduate degree program in human rights.
During the Short-Stay Study Abroad Program 2017, I worked under the supervision of Prof. Tam
Nguyen at “Vision and Mixed Reality Lab” housed in the Department of Computer Science,
University of Dayton. Prof. Tam Nguyen is an Assistant Professor in the Department of Computer
Science, University of Dayton. His research works about visual saliency models have been
published in premier venues, i.e., ACM Multimedia, European Conference on Computer Vision,
International Joint Conference on Artificial Intelligence, International Journal on Computer Vision,
IEEE Transactions on Multimedia, IEEE Transactions on Circuits and Systems for Video
Technology, Neurocomputing, and Multimedia Tools and Applications Journal. My Ph.D. topic
has some significant overlaps with his research (i.e., computer vision and deep learning). Therefore,
I joined a research project about camouflaged object segmentation at the University of Dayton
under his guidance. I firmly believe the Short-Stay Study Abroad Program is a good opportunity
for me to learn more, to see more, to broaden my vision and to improve myself.
Research Topic: Camouflaged Object Segmentation
Camouflaged objects attempt to conceal their texture into the background and discriminating them
from the background is hard even to us human beings. The main objective of this paper is to explore
the camouflaged object segmentation problem, namely, segmenting the camouflaged object(s) for
a given image. This problem has not been well studied in spite of a wide range of potential
Present year: 2017 2017 Doctoral course
applications including discovering and preserving wildlife animals, surveillance systems, human
search-and-rescue missions in natural disasters such as earthquake, flood, or hurricane.
The objective of my Ph.D. research is to utilize deep learning for salient object segmentation.
Salient object predictors aim to detect and segment salient objects in images. Though “salient” is
opposed to “camouflage”, techniques developed for salient object segmentation can be used for
camouflaged object segmentation. This is because the two tasks target at highlighting some image
regions with certain characteristics. Thus, this research project totally benefits for my Ph.D. study
at NII, SOKENDAI. That is the reason that I applied to the Short-Stay Study Abroad Program in
FY 2017.
Research Schedule
While joining the Vision and Mixed Reality Lab at the University of Dayton, I worked on the
project of camouflaged object segmentation using deep learning models. My work at the Vision
and Mixed Reality Lab focuses on building a new dataset and conducting experiments. The plan
details are as follow:
From 9/18 to 9/27, I collected 5000 images from the Internet and filtered out unsuitable
images, kept 2500 images. This work was supported by two students from the University
of Dayton.
From 9/30 to 10/7, I and two students from the University of Dayton manually annotated
the ground-truth label for the dataset.
From 10/8 to 10/12, I trained the network and conducted necessary experiments.
After that, I returned to Japan and wrote a research paper at NII, SOKENDAI. We finished the
paper and will submit to CVPR 2018 (the deadline is 11/15/2017). All discussion and paper writing
are performed in English.
Research Results
In this project, we provided a new image dataset of camouflaged object segmentation where each
image is manually annotated with pixel wise ground-truth. We also proposed a dual stream network
(DSN) for camouflaged object segmentation where the classification stream and the segmentation
stream are effectively combined to achieve good performance. Extensive experiments conducted
on the newly built dataset demonstrate the superiority of the proposed framework. In addition, our
method is computationally efficient, running in real time. We believe that our provided dataset will
promote work to further explore this interesting yet challenging problem.
Figure 1. Example annotations of our collected dataset with ground-truth segmentation masks overlaid.
Figure 2. Visual comparison of our method against the state-of-the-art methods. From left to right, input image and ground-truth
are followed by outputs obtained using our method (DSN) and the other methods. The first nine rows are images with camouflaged
object. Our method surrounded with red rectangles achieves the best results.
Expenses
The expenses are used as below:
Visa and insurance: 33,000 JPY.
Domestic travel expenses: 2,960 JPY in Japan and 140 USD in US.
Round-trip international airfares: 172,960 JPY.
Accommodation: 161,897 JPY.
Living expenses: 700 USD.
Opinions
I would like to thank SOKENDAI and the Short-Stay Study Abroad Program. The trip to
University of Dayton sponsored by the Short-Stay Study Abroad Program is a good opportunity
for me to learn more, to see more, to broaden my vision and to improve myself. In addition, I made
new friends, worked with new professors, as well as had more connection for further research in
the future.
Some Photos
Figure 3. Vision and Mixed Reality Lab, University of Dayton.